我正在使用 cuda 7.5 版cufft来执行一些 FFT 和逆 FFT。使用cufftExecC2R(.,.)函数执行逆 FFT 时出现问题。
实际上，当我在中使用 abatch_size = 1时，cufftPlan1d(,)我得到了正确的结果。但是，当我增加批量大小时，结果不正确。
我正在粘贴一个示例最小代码来说明这一点。请忽略代码的肮脏，因为我刚刚快速创建了它。
#include <cufft.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#include <ctime>
#include <iostream>
typedef float2 Complex;
void iTest(int argc, char** argv);
#define SIGNAL_SIZE 9
#define BATCH_SIZE 2
int main(int argc, char** argv) {
iTest(argc, argv);
return 0;
}
void iProcess(Complex *x, double *y, size_t n) {
cufftComplex *deviceData;
cudaMalloc(reinterpret_cast<void**>(&deviceData),
SIGNAL_SIZE * BATCH_SIZE * sizeof(cufftComplex));
cudaMemcpy(deviceData, x, SIGNAL_SIZE * sizeof(cufftComplex) * BATCH_SIZE,
cudaMemcpyHostToDevice);
cufftResult cufftStatus;
cufftHandle handle;
cufftStatus = cufftPlan1d(&handle, SIGNAL_SIZE, CUFFT_C2C, BATCH_SIZE);
if (cufftStatus != cudaSuccess) {
printf("cufftPlan1d failed!");
}
cufftComplex *d_complex;
cudaMalloc(reinterpret_cast<void**>(&d_complex),
sizeof(cufftComplex) * SIGNAL_SIZE * BATCH_SIZE);
cufftStatus = cufftExecC2C(handle, deviceData, d_complex, CUFFT_FORWARD);
if (cufftStatus != cudaSuccess) {
printf("cufftExecR2C failed!");
}
cufftComplex *hostOutputData = (cufftComplex*)malloc(
(SIGNAL_SIZE) * BATCH_SIZE * sizeof(cufftComplex));
cudaMemcpy(hostOutputData, d_complex,
SIGNAL_SIZE * sizeof(cufftComplex) * BATCH_SIZE,
cudaMemcpyDeviceToHost);
std::cout << "\nPrinting COMPLEX" << "\n";
for (int j = 0; j < (SIGNAL_SIZE) * BATCH_SIZE; j++)
printf("%i \t %f \t %f\n", j, hostOutputData[j].x, hostOutputData[j].y);
//! convert complex to real
cufftHandle c2r_handle;
cufftStatus = cufftPlan1d(&c2r_handle, SIGNAL_SIZE, CUFFT_C2R, BATCH_SIZE);
if (cufftStatus != cudaSuccess) {
printf("cufftPlan1d failed!");
}
cufftReal *d_odata;
cudaMalloc(reinterpret_cast<void**>(&d_odata),
sizeof(cufftReal) * SIGNAL_SIZE * BATCH_SIZE);
cufftStatus = cufftExecC2R(c2r_handle, d_complex, d_odata);
cufftReal odata[SIGNAL_SIZE * BATCH_SIZE];
cudaMemcpy(odata, d_odata, sizeof(cufftReal) * SIGNAL_SIZE * BATCH_SIZE,
cudaMemcpyDeviceToHost);
std::cout << "\nPrinting REAL" << "\n";
for (int i = 0; i < SIGNAL_SIZE * BATCH_SIZE; i++) {
std::cout << i << " \t" << odata[i]/(SIGNAL_SIZE) << "\n";
}
cufftDestroy(handle);
cudaFree(deviceData);
}
void iTest(int argc, char** argv) {
Complex* h_signal = reinterpret_cast<Complex*>(
malloc(sizeof(Complex) * SIGNAL_SIZE * BATCH_SIZE));
std::cout << "\nPrinting INPUT" << "\n";
for (unsigned int i = 0; i < SIGNAL_SIZE * BATCH_SIZE; ++i) {
h_signal[i].x = rand() / static_cast<float>(RAND_MAX);
h_signal[i].y = 0;
std::cout << i << "\t" << h_signal[i].x << "\n";
}
std::cout << "\n";
double y[SIGNAL_SIZE * BATCH_SIZE];
iProcess(h_signal, y, 1);
}
我无法找出我的代码中的错误在哪里以及我缺少哪些信息。
使用时的示例输出BATCH_SIZE = 1
使用时的示例输出BATCH_SIZE = 2
